What LLM-based software engineering actually looks like today

AI and software development: where do we stand?

The presenter opens by noting that AI now dominates nearly every tech headline. But despite the excitement, he stresses that real engineers report a much more mixed and nuanced experience when it comes to AI-assisted software development.

"Microsoft's CEO claims that 30% of the code is now written by AI. What does that even mean? Is that actually impressive?"

Executives celebrate AI loudly, but the field is still full of cautionary stories. One startup engineer paid $500 per month for an agent called Devon, only to spend another $700 cleaning up a bug it introduced.

"We all know AI isn't perfect. This story is just one example."

After Microsoft's Build conference, a failed attempt to deploy Copilot inside a .NET codebase trended on Reddit.

"I admire Microsoft for sharing that failure, but it underscores how real limits still remain."

How AI tools are actually used

Drawing from conversations with engineers, the presenter groups real-world usage stories into four buckets.

1. AI tool startups

• Anthropic

  • As soon as Cloud Code rolled out internally, engineers adopted it day-to-day.
  • People even say 90% of their Cloud Code is written by Cloud Code today.
  • Usage jumped 40% on launch day and 160% within a month.
  • They open-sourced MCP (Model Context Protocol) so it can interoperate with many tools.
  • "You can now ask it how many people used a promotion code, and it emits the SQL for you."

• Windinsurf

  • "Roughly 95% of our code goes through our AI agent or workspace tabs."

• Cursor

  • "About 40–50% of the code is AI-generated. Some parts work brilliantly; others still lag."
  • They still aim for 100% automation but admit the limitations are real.

2. Big tech

• Google

  • Every environment is custom: Cider IDE, homegrown code-review dashboards, and internal tools.
  • "AI is everywhere at Google now—LLMs power completion, chat inside Cider, automated reviews."
  • "A year ago this barely existed; now it's moving fast."
  • "They're rolling it out slowly so engineers can build trust."
  • "SRE friends say they are strengthening infrastructure, deployment pipelines, reviews, and feature flags because they expect production code lines to grow tenfold."

• Amazon

  • "Almost every Amazon developer uses Q Developer Pro for AWS code."
  • "It was mediocre six months ago; now it's quite good."
  • "Nearly every internal tool and website speaks MCP—automation is everywhere."
  • "API-first culture made MCP adoption feel natural."

3. Other startups

• incident.io

  • "We use AI to boost productivity and share tips inside Slack."
  • "Well-defined tickets can safely be handed to an agent."
  • "Cloud Code is now a regular part of the team's toolkit."

• Anonymous biotech AI startup

  • "We tried many LLMs, but manual coding is still faster for us."
  • "We build completely new software, so AI hasn't aligned well yet."

4. Independent engineers

• Armin Ronacher (Flask creator)

  • "Six months ago I would have hired a human lead over a virtual agent. Now I think differently."
  • "Cloud Code works so well that hallucinations are less worrying when the tool actually executes commands."

• Peter Steinberger (PSPDFKit creator)

  • "I haven't been this excited about technology in years. It feels like a turning point."
  • "Language or framework barely matters anymore; you can swap them easily."
  • "Engineers are staying up all night from excitement."

• Bri Brigita (Thoughtworks)

  • "LLMs span every abstraction layer. It's not just another layer; it's a lateral move across the entire stack."

• Simon Willison (Django co-creator)

  • "Coding agents actually work now—models have rapidly improved in the last six months."

Reality check: the unresolved questions

The presenter lists four lingering questions.

1. Why are CEOs and founders more enthusiastic than engineers?

   • "Warp's founder says senior engineers barely use AI, while founders and PMs evangelize it."
   • "CEOs publicly gush—why the enthusiasm gap?"

2. Is AI tooling mainstream yet?

   • "About 60–70% of you use AI tools weekly."
   • DX's survey of 38,000 engineers shows roughly half of organizations use AI tools weekly; high-performing companies hit 60%.
   • "That would have been unimaginable three years ago, but it still isn't universal."

3. How much time does it actually save?

   • "Peter says output is 10–20x, yet DX reports about 3–5 hours saved per week—four hours on average."
   • "On a 40-hour week that's not a tenfold gain. What do you do with those freed hours?"

4. Why is it great for individuals but still limited at scale?

   • "Everyone agrees the tools serve solo developers well, but organization-wide deployments are still a work in progress."

Looking ahead

He notes that experienced engineers are adapting successfully to AI tools.

"It feels like development is undergoing a staged transformation."

Then he cites the legends.

• Martin Fowler

  "LLMs will trigger a shift similar to the one from assembly to high-level languages—except this time it's a nondeterministic revolution."

• Kent Beck (52-year veteran)

  "After 52 years of programming, I'm having the most fun." "Like microprocessors, the internet, or smartphones, LLMs will reshape tooling." "Tasks once too expensive or complex are now cheap and easy. It's time to experiment."

Summary

• AI developer tools are in heavy use, but not yet perfect for every organization.
• Individual and seasoned engineers are realizing productivity wins and rekindling their enthusiasm.
• Organization-wide deployments and true productivity maximization remain challenges.
• We should keep experimenting and reevaluate what is cheap versus expensive.

"Change is happening—experiment more and keep trying. Thank you!"

🎉